FaithNet: A Generative Framework in Human Mentalizing

In this paper, we first review some of the innovations in modeling mentalizing.Broadly, this involves building models of computing World Model and Theory of Mind(ToM). A simple framework, FaithNet, is then presented with concepts like persistence, continuity, cooperation and preference represented as faith rules.FaithNet defines a generative model that can sample faith rules. Our FaithNet utilize a general-purpose conditioning mechanism based on cross-attention, offering computations that best explain observed real-world events under a Bayesian criterion.